Skip to content
JobShout
Insights
Blogaiawsazuredata-residencyllmprivate-ai

Private AI Hosting in the UK: AWS Bedrock vs Microsoft Azure vs Your Own GPUs

Compare AWS Bedrock, Microsoft Azure, UK AI providers and your own GPUs for customer financial data: UK and EU data residency, cost per call and when a GPU pays off.

Workstation5 min read

ShareLinkedIn (opens share dialog)X (opens share dialog)Email (opens share dialog)

If your product sends customer financial data to an AI model, two questions decide almost everything: where is that data processed, and what does each call cost? We compared AWS Bedrock, Microsoft Azure, UK and EU AI providers and running your own GPUs for a typical UK finance workload. The short answer: pay-per-token AI costs nothing when idle, Bedrock is the best fit for most UK firms, and the first decision is not technical at all.

In one sentence. Decide whether customer data must stay in the UK or may be processed in the EU; then use Amazon Nova in Bedrock's London region (UK only) or Claude Haiku through Bedrock's EU profile (EU), cut volume with rules and prompt caching, and test quality on your own data before you compare prices.

The workload

Our worked example is a UK bookkeeping or tax application. Most of its AI spend is classifying bank transactions: one call per transaction, with about 3,600 tokens of instructions and context going in and about 250 tokens coming out. It also reads scanned receipts, which needs a model that understands images, and runs a chat assistant for customers.

The rule the business set: customer financial data must not go to public consumer AI services. It must be processed in the UK, or at least the EU, with no training on it and no long retention.

Myth: switch AI off at night to save money

Both Bedrock and Azure's standard deployments charge per token. When nothing is being classified, they cost nothing. Switching them off saves money only for services billed by the hour: reserved capacity (Bedrock Provisioned Throughput, Azure PTU) or a rented GPU. On a pay-per-token service, the on/off switch decides where your data goes, not what you pay.

Start with where the data may go

Workflow: choosing a private AI route from your data-residency rule. Start with one question: must customer data be processed only in the UK? If yes, the options are Amazon Nova Micro or Nova Lite on Amazon Bedrock in London, Civo relaxAI, or your own GPU server. If the EU is acceptable, Claude Haiku on Bedrock with the EU inference profile is the default, with Scaleway in France or a dedicated EU endpoint as alternatives. In either case avoid routes that cannot pin a region: Claude on Microsoft Foundry has no EU or UK processing option yet, and Azure OpenAI in UK South needs Microsoft approval. Every route then goes through the same quality test before price is compared.
Decide the residency rule first. UK-only points to Amazon Nova in Bedrock London, Civo relaxAI or your own GPUs; EU-allowed adds Claude Haiku through Bedrock's EU profile and EU providers. Routes that cannot pin a region are out for regulated data.
  • AWS Bedrock runs Amazon's Nova Micro and Nova Lite models in-region in London, so data stays in the UK. Claude Haiku is available through an EU inference profile, processed across EU regions but not pinned to the UK.
  • Microsoft Azure offers Claude on Microsoft Foundry, but at the time of writing with no EU or UK processing option. Its UK South region runs OpenAI models such as GPT-5 mini, but customers report it needs Microsoft's approval and capacity is not guaranteed.
  • UK and EU providers such as Civo's relaxAI (UK data centres) and Scaleway (France) offer open models through an OpenAI-compatible interface.
  • Your own GPU server keeps data entirely under your control.

What each route costs

Approximate monthly cost per 100,000 classifications by route. Illustrative cost per month for 100,000 calls of about 3,600 input and 250 output tokens each, at November 2026 list prices. Amazon Nova Micro in London about 23 dollars; Nova Lite in London about 39; Claude Haiku on Bedrock EU with prompt caching about 24; Nebius Qwen3-32B about 44; Claude Haiku on Microsoft Foundry about 49 but with no EU or UK option; Claude Haiku on Bedrock EU about 53; Civo relaxAI about 66; Scaleway Mistral Small about 72; GPT-5 mini on Azure OpenAI UK South about 156 plus reasoning tokens. A rented UK GPU costs about 942 dollars a month whatever the volume, and your own server costs nothing extra per call. Bars are coloured by where data is processed: UK, EU, or not pinned.
Approximate monthly cost per 100,000 classifications at November 2026 list prices. Divide by ten for 10,000 a month, multiply by ten for a million. Re-check every price in the provider's own calculator.

For 100,000 transactions a month, the per-token routes range from about $23 (Nova Micro in London) to about $53 (Claude Haiku through Bedrock's EU profile), with UK-hosted GPT-5 mini on Azure around $156 before its hidden reasoning tokens. A rented UK GPU costs about $942 a month whatever the volume.

Price is only half the story. The cheapest model is only cheap if it classifies correctly. Nova and the open models are untested on most firms' data, while Claude Haiku is the strongest of the low-cost options and reads images. That is why the next step is a quality test, not a purchase order.

Cut the bill before you pick the model

  1. Rules first. Obvious payees and merchant codes need no AI at all.
  2. Remember answers. The same supermarket every week does not need a fresh AI call each time.
  3. Prompt caching. Every call repeats the same category list and rules. Cached, that part costs about a tenth of the normal price on Claude; for this workload that roughly halves the bill.
  4. Batch overnight work. Reclassifying old transactions is half price as batch work on Bedrock.

Together these can cut a $53 monthly bill to under $10 for the same 100,000 transactions, depending on how repetitive your customers' spending is.

When does your own GPU pay off?

When does a dedicated GPU beat pay-per-token?. A chart of monthly cost against monthly classifications from zero to two million. Claude Haiku on Bedrock EU rises in a straight line, about 53 dollars per 100,000, reaching about 1,067 dollars at two million. Amazon Nova Micro in London rises more slowly, about 450 dollars at two million. A rented UK GPU is a flat line at about 942 dollars a month. The GPU only matches Haiku at about 1.8 million classifications a month and never matches Nova Micro in this range. Pay-per-token routes cost nothing when idle.
A rented UK GPU at about $942 a month only matches Claude Haiku at roughly 1.8 million classifications a month, and does not match Nova Micro within two million. Hardware you already own, or a GPU shared by several AI workloads, changes the sum.

For most SMEs, pay-per-token wins on cost until volumes are very high. A dedicated GPU makes sense when you already own the hardware, when data must never leave your control, or when one server can carry several AI workloads at once: classification, document reading, search and internal assistants.

Our recommendation

  1. Decide the data rule with whoever owns data protection: UK only, or EU allowed.
  2. Test quality before price on about 200 transactions with known right answers, across every route the rule allows.
  3. Default to AWS Bedrock: Nova in London if UK-only, Claude Haiku through the EU profile if EU is allowed. Add prompt caching and a monthly budget alert from day one.
  4. Keep Azure for cases where you specifically need OpenAI models processed in the UK, or Microsoft contracts matter to your customers.
  5. Keep a UK fallback, such as your own GPU or a UK provider, so one provider's outage or model retirement does not stop the business.

The full working, with the cost tables at three volumes, Bedrock and Azure residency details, IAM and code examples, the request pipeline and the evaluation harness, is in our technote: Private AI for financial data: residency, routing and cost on Bedrock, Azure and your own GPUs.

How Workstation helps

We help UK organisations run AI on sensitive data: choosing residency-aware routes on Bedrock, Azure and UK providers, building the quality tests, cutting cost with caching and rules, and supplying or hosting GPU servers when running your own model makes sense. See our GPU private cloud guide, compare leasing and buying at gpuleasing.co.uk, or talk to us.

Prices are approximate November 2026 list prices in US dollars (£1 = $1.30, €1 = $1.15), from provider pages and third-party trackers; check each provider's calculator before committing. Figures are illustrative for the workload described, not a quote.

Acknowledgement of rights: AWS, Amazon Bedrock and Amazon Nova are trademarks of Amazon.com, Inc. or its affiliates. Microsoft, Azure and Microsoft Foundry are trademarks of the Microsoft group of companies. Claude is a trademark of Anthropic, PBC. OpenAI and GPT are trademarks of OpenAI. Other names are the property of their respective owners.